Anthropic adds invisible watermark to Claude AI text output

Every piece of text that Claude generates is about to carry a hidden signature. Anthropic announced Monday that Claude models released on or after August 2, 2026, will embed an imperceptible, machine-readable watermark directly into all generated text, invisible to human readers but detectable by the right tools.

The watermark leaves the meaning and readability of the text completely intact. It survives copy-paste operations and holds up through at least moderate editing, which is exactly the point: the goal is to keep the signal alive even after the text has moved around the internet a few times.

How it actually works

The watermark is woven into the statistical patterns of the text, the subtle choices in word selection and phrasing that a language model makes billions of times per output. Humans cannot see it. Machines, with the right detection tool, can.

Anthropic will also implement digitally signed provenance metadata built on the C2PA standard, a technical framework already used to verify the origin of images and video.

Public detection tools will be released alongside the feature, though Anthropic is being measured about what they can actually promise. The watermark’s reliability degrades under heavy editing, and short outputs may not carry a strong enough signal to detect reliably.

The company also said it plans to retroactively bring the watermarking capability to older Claude versions during the transition period leading up to August 2026, which gives developers and enterprise API users time to adjust.

The regulatory backdrop

The EU AI Act includes specific transparency obligations under Article 50. Those rules require providers of general-purpose AI systems to ensure that content generated by their models is marked in a way that is machine-detectable. Anthropic is treating the watermark as a global standard rather than a Europe-only compliance patch, meaning Claude users worldwide will be subject to the same marking regime.

Google’s DeepMind team developed a tool called SynthID that embeds watermarks into AI-generated text, images, and audio.

What this means going forward

The caveat that Anthropic is being careful to flag: watermarks are not proof of absence. Text that doesn’t carry a watermark is not necessarily human-written. Someone could generate content with a model that doesn’t watermark, or strip a watermark through aggressive enough editing. The signal tells you something is Claude-generated when it is present. It cannot tell you everything is human-generated when it is absent.

Disclosure: This article was edited by Editorial Team. For more information on how we create and review content, see our Editorial Policy.